ML Fundamentals MCQs in Placement Tests: 2026 Guide

ML MCQs in Indian placement tests are definition-level: which algorithm solves which problem, not how to derive its gradient.

That distinction matters a lot for prep. You don’t need calculus or coding for the ML section of TCS NQT, Infosys, or Cognizant drives. You need to know the vocabulary, the use cases, and the key concepts well enough to pick the right option in under 60 seconds.

Here is what gets tested in 2026 and how to prepare for it.

Where ML MCQs appear in 2026 placement drives

Not every company tests ML in the same way or at the same depth. The table below maps what you can expect across common campus drives:

Company / Test	ML coverage	Depth
TCS NQT (standard)	Integrated into advanced reasoning section	Definition-level: learning types, algorithm use cases
TCS AI Career	Dedicated ML assessment	Python + ML concepts, algorithm matching, library awareness
Infosys online test	Technical MCQ section	Awareness-level: supervised vs. unsupervised, basic definitions
Cognizant GenC / GenC Next	AI/ML awareness questions	Type identification, metric definitions
Wipro NLTH	Minimal in initial screen	More common in later technical assessment rounds

The pattern across companies: initial screening tests lean on definitional MCQs. If you can state what logistic regression does, name a step in k-means, or explain what precision measures, you are prepared for what the test actually checks.

TCS NQT’s standard exam runs Part A (65 questions in 75 minutes covering numerical, verbal, and reasoning ability) and Part B, which includes advanced reasoning and coding. ML and AI awareness questions sit inside the advanced section, not as a standalone block. TCS AI Career is a separate programme with a dedicated technical assessment that goes deeper into Python and ML concepts.

High-frequency ML topics across placement tests

Eight topic clusters account for most ML MCQs that appear in campus drives. The table lists them in rough order of question frequency:

Topic cluster	What’s tested	What’s NOT tested in MCQs
Learning types	Supervised, unsupervised, reinforcement: definitions and examples	Multi-task learning, federated learning
Classification metrics	Confusion matrix, precision, recall, F1, accuracy	Step-by-step AUC-ROC calculation
Core algorithms	Logistic regression, k-NN, k-means, decision trees, Naïve Bayes, SVM concept	Deriving the SVM kernel function
Overfitting and regularisation	Overfitting definition, bias-variance tradeoff, L1 vs. L2 concept	Proof of bias-variance decomposition
Ensemble methods	Random forest concept, bagging vs. boosting distinction	Gradient boosting math
Neural networks	Neuron structure, activation functions (ReLU, sigmoid, tanh), backpropagation idea	Deriving backpropagation equations
Data preprocessing	Feature scaling, one-hot encoding, missing value handling	Detailed imputation method tradeoffs
Dimensionality reduction	PCA purpose and intuition	Eigenvalue computation

The working rule for each topic: know the definition, one real-world use case, and what makes it different from the most similar concept. That covers the bulk of what placement MCQs actually ask.

Sample ML MCQs: questions, answers, and the reasoning

These 12 questions cover the topic clusters above. Each answer includes a brief explanation because placement MCQs often re-use the same logic in different phrasings.

Learning types and definitions

• Q1: Which type of machine learning uses labeled training data to learn a mapping from inputs to outputs?

• Options: Supervised learning / Unsupervised learning / Reinforcement learning / Transfer learning

• Answer: Supervised learning. The defining feature is labeled input-output pairs during training. Unsupervised learning finds structure without labels; reinforcement learning learns through reward feedback.

• Q2: A model groups customer purchase records into segments with no predefined category labels. Which learning type does this represent?

• Options: Supervised / Unsupervised / Reinforcement / Semi-supervised

• Answer: Unsupervised learning. No labels are used. k-means and hierarchical clustering are the canonical examples tested in campus MCQs.

Algorithm use cases

• Q3: Despite having “regression” in its name, logistic regression is primarily used for which task?

• Options: Predicting continuous values / Clustering / Classification / Dimensionality reduction

• Answer: Classification. The sigmoid function converts the model’s raw output into a probability between 0 and 1, which is thresholded to assign a class label.

• Q4: What makes k-NN a “lazy learner”?

• Options: It uses a simplified model / It ignores irrelevant features / It stores all training data and defers computation to prediction time / It learns slowly on large datasets

• Answer: It stores all training data and defers computation to prediction time. No explicit model is built during training. This makes training fast but prediction slower on large datasets.

• Q5: Which algorithm is used for unsupervised clustering of data into k groups?

• Options: Logistic regression / k-means / Naïve Bayes / SVM

• Answer: k-means. The other three are supervised algorithms that require labeled training data.

• Q6: How does random forest reduce overfitting compared to a single decision tree?

• Options: By using a deeper tree / By training many trees on random subsets and averaging predictions / By using fewer features / By applying L1 regularisation

• Answer: By training many trees on random subsets of data and averaging their predictions. This ensemble averaging smooths out individual tree errors, reducing variance without sacrificing much bias.

Evaluation metrics

• Q7: A classifier reports: TP = 80, FP = 20, FN = 10, TN = 90. What is its precision?

• Step 1: Precision = TP / (TP + FP)

• Step 2: 80 / (80 + 20) = 80 / 100 = 0.80

• Answer: 0.80. The Wikipedia confusion matrix article has a complete breakdown of all four terms if the notation is new.

• Q8: When is recall more important than precision?

• Options: When false positives are costly / When false negatives are costly / When the dataset is balanced / When the model is simple

• Answer: When false negatives are costly. Medical diagnosis is the standard example: missing a disease (false negative) is far worse than a false alarm (false positive). The scikit-learn evaluation guide covers precision, recall, F1, and macro vs. micro averaging in one reference.

• Q9: Which metric is the harmonic mean of precision and recall?

• Options: Accuracy / F1-score / AUC-ROC / Matthews Correlation Coefficient

• Answer: F1-score. The harmonic mean penalises extreme values. A high F1 requires both precision and recall to be reasonably high, not just one of them.

Overfitting and neural network basics

• Q10: A model scores 99% on training data but 61% on unseen test data. What does this indicate?

• Answer: Overfitting. The model memorised training-data noise rather than the underlying pattern. The overfitting and regularisation guide covers L1 vs. L2 remedies in detail, a common follow-up MCQ topic.

• Q11: High bias combined with low variance in a model indicates:

• Options: Overfitting / Underfitting / Good generalisation / Noisy training data

• Answer: Underfitting. The model is too simple to capture the underlying pattern. The opposite situation, low bias with high variance, indicates overfitting.

• Q12: Which activation function is most commonly used in hidden layers of deep neural networks?

• Options: Sigmoid / Tanh / ReLU / Softmax

• Answer: ReLU (Rectified Linear Unit). It avoids the vanishing gradient problem that sigmoid and tanh face in deeper networks and is computationally inexpensive to evaluate.

How to sequence your ML prep

Three weeks of prep alongside aptitude and coding work is enough for most standard placement drives. Here is the priority order:

Week 1 (highest ROI): Evaluation metrics and learning types

• Confusion matrix: TP, FP, TN, FN — practice calculating precision, recall, F1, and accuracy from raw numbers. These show up across TCS, Infosys, and Cognizant tests.
• Learning types: supervised, unsupervised, reinforcement. One concrete example of each.
• Overfitting vs. underfitting: the bias-variance tradeoff at a conceptual level. The bias-variance tradeoff guide walks through the intuition and gives you the framing for interview answers, not just MCQs.

Week 2 (core algorithms): The main six

• Logistic regression, k-means, decision trees, k-NN, SVM (margin concept), random forest (bagging concept).
• For each: what problem it solves, one example, and what distinguishes it from the most similar algorithm.
• Skip the math derivations. Know the intuition and the tradeoffs.

Week 3 (for AI-specific roles): Neural nets and tools

• Activation functions: ReLU, sigmoid, tanh, softmax, and when each is used.
• Data preprocessing: feature scaling, one-hot encoding, handling missing values.
• Python library roles: what scikit-learn, NumPy, and pandas each do at a high level.
• If your drive has a coding component beyond MCQs, the Python coding questions guide for AI/ML interviews covers the library and syntax questions that come up when tests move from multiple-choice to code.

The ML MCQ section of placement tests rewards students who know the vocabulary precisely. “Logistic regression is for classification,” “k-means is unsupervised,” and “high bias means underfitting” are the types of distinctions that determine correct answers. Equations are not the test.

For a picture of the full AI skill stack beyond what placement MCQs check, the 2026 AI roadmap for Indian engineering students maps everything from fundamentals to deployment. That is the useful next read once you have cleared the MCQ screen. If your drive also tests pandas and NumPy operations, the pandas and NumPy interview questions guide covers the DataFrame and array questions that accompany the conceptual MCQs.